← The AI Policy Podcast
The AI Policy Podcast
Senate Strikes AI Law Moratorium, Courts Rule on Copyright Cases, and Congress Talks AGI
00:33:52
Loading summary
Transcript33 lines
- [00:00]
A
Foreign.
- [00:10]
B
Welcome back to the AI Policy podcast. Me and Greg are ready to rock. Today we'll start with the Senate's vote to remove the moratorium on state AI laws from the reconciliation bill this literally just passed. We're talking at 2pm on July 1st. Will also cover major updates on AI copyright rulings and the House Select Committee on China's recent AI hearing, plus some new developments surrounding Deep Seq. But first, Greg, we got to go to the news. Moratorium on state AI laws. Just a few hours for this conversation. Senate voted 99 to 1 to strip the state AI moratorium from the so called big beautiful bill. If you've been following along since our June 4 episode, you'll know that this moratorium has been through several revisions before the Senate ultimately cut it. So, Greg, what exactly was the original provision the House passed and how did the senators reshape it?
- [01:05]
A
Well, first I'd like to take a victory lap because on this podcast we predicted that it was not going to be a part of the final big beautiful bill.
- [01:14]
B
That's right.
- [01:15]
A
And would not pass. And that is indeed what has come to pass. So. So victory lap taken. Yeah. This idea was to basically block not all, but certain types of state level AI regulation, specifically the kinds of regulation that would impose sort of preemptory actions that AI companies have to do before they release their AI products. So it didn't get rid of, you know, criminal penalties for people who use AI to do bad stuff, but it did prevent most categories of new, new requirements that were AI specific that AI companies had to do before they, you know, released the model or something like that. And that turned out to be exceptionally controversial. It passed in the House, but some folks who supported it in the House, voted for it in the House, later said that they didn't understand what they were voting for and said that they regretted their vote and would not vote for it again. So those were the conditions under which it went to the Senate where it encountered, you know, new categories of opposition, most notably Senator Marsha Blackburn of Tennessee. Tennessee has a very strong recording music industry in Nashville, very strong in country music and some other genres as well. And so she was concerned that this preemptory move would take down laws in Tennessee. But that also built some momentum among other senators who were thinking about this. There was genuinely bipartisan opposition to this. As we mentioned on the last podcast, a letter co signed by 40 states attorneys general opposed this. So it looked like actually there was some positive momentum in this. For one thing, the Senate parliamentarian ruled that it could be included as part of the big beautiful bill, which I was surprised by, given that it was not clear at all to me how it related to revenue stuff, which is usually the pretty strict interpretation of the rule for including something in a reconciliation package. So it survived contact with the parliamentarian, but it did not survive contact with opposition among Senate Republicans. And even during a period where Senator Ted Cruz, who has been a supporter of this moratorium, and Senator Blackburn reached something of a compromise, that compromise later fell apart over a disagreement about, you know, whether that compromise amendment language was actually going to have the intended purpose or was actually executable. So Blackburn withdrew her support for the compromise. And then the. The bill was voted out of the Senate by, as you said, 99 to 1. So when I think about, like, who is comprising that 99 to 1, that's basically everybody who opposes this to begin with and hopes that it doesn't happen, and everybody else who agrees that they don't want to torpedo the big beautiful bill over this issue for which there is not a majority of support. So even people who support this moratorium were voting to strip it out of the big beautiful bill. And that's where we are right now. It's still a future where there are proponents who want to have some kind of federal action, but the states are not going to be blocked in going ahead with their regulation. And there is a flurry of activity at the state level. Literally more than a thousand bills have been introduced. And that gets to the biggest concern of the AI companies, which I think is a relatively strong argument for why regulation should be at the federal level. They said that, you know, this patchwork of different regulations among the different 50 states is going to be extremely difficult for them to implement and is going to provide a barrier to their innovation, barrier to AI adoption, something that the United States should not want. But there's a rebuttal to that argument, which is the patchwork, I guess I should say. First, the companies have been saying pretty consistently that they do not oppose AI regulation around, you know, consumer protection or AI safety or countering deepfake pornography, all of those kinds of things. They're open to regulation. They just want it to be at the federal level. And that's why I thought a quote from Miles Brundage, who is a friend of the podcast, asked he he said on X on June 30, AI should be regulated federally is why the moratorium is bad, not why it's good. Blocking state regulation without doing anything federally is like taking a strong pain Reliever that makes you pass out and forget to treat the real injury. The patchwork is the incentive. And so that's why I say actually voting this down might increase the chance of some kind of moratorium going into effect in the future if and only if that moratorium is accompanied by a federal AI regulatory framework. Something that, you know, the government can say with a straight face, we don't need state action, because here is the federal action. So that's where I think we are now.
- [06:35]
B
So, Greg, is. Are states going to our companies going to start going state shopping now because of this or, you know.
- [06:43]
A
Yeah. So I mean, ultimately the states do not have the power of interstate commerce. And all of these companies are providing their services, you know, over the border. Right. Most of them have data centers in California or Virginia or wherever, and then they're delivering online services in the form of these chatbot agents that go across state borders. However, the states do have the right to police, you know, what is going on within their borders, the types of transactions, the types of services that are being provided. And so, you know, regulatory district shopping actually doesn't solve the problem from the perspective of AI companies because they're looking to have clarity on their regulatory burden for the users that they're trying to serve. And they want to serve users in all 50 states.
- [07:36]
B
All right, let's change up here a bit to something else that's legal. Copyright court rulings. Meta and Anthropic both recently won key copyright cases over the use of books to train their AI models. This is a fascinating case. Before we get into the details of these rulings, who were the plaintiffs in these cases and what were they each claiming?
- [07:58]
A
Yeah. So I think it's worth pointing out that both of these court orders, which came from different US District courts in California, are pretty interesting documents. You can see here, you know, I've got the Meta one and I highlighted basically the entire first page. And the reason for that is that there's a lot of meaty stuff in both of these court orders talking about the practices of these companies. Just to give you, you know, one flavor. Both Anthropic and Meta did download millions of pirated books, books for use in AI training data. In the Anthropic decision, it, you know, references some internal emails where Anthropic talks about books as an incredibly high quality source of training data. You know, so if you think about how do I make my AI model smarter, do I want it reading, you know, the 10,000th random forum post on some random Internet site or Do I want it reading, you know, the best book of the year on XYZ topics topic. It turns out, unsurprisingly, right, that you get a bigger boost in performance by having higher quality training data. And all the AI companies have assessed that books are utterly critical. Right. To give the performance that they're looking to give their consumers.
- [09:21]
B
So let me stop, let me stop there for one second. What do you mean by pirated books? Books that they didn't pay for?
- [09:26]
A
They didn't pay for. Literally. They went to sites where these are known piracy websites that have downloadable repositories of millions of books. Everybody in all of these cases acknowledges that piracy is illegal. In Anthropic's case, I mean, they were obsessed with speed. So was Meta. Right. It's not like they don't have the money to pay for the books. These are both incredibly well financed corporations. The variable in the equation that mattered to them was speed. They didn't want to negotiate 1 million different, you know, licenses to try and do this. So in Anthropic's case, according to what?
- [10:06]
B
So they just stole it?
- [10:07]
A
Yeah, so they just stole it. And again, like they're basically admitting that. So in Anthropic's case, they started with the pirated books because they were worried about, you know, just how slow it was going to be to go acquire all the rights. But then, and this is coming out in internal emails which were released as part of the discovery process in these cases, they ultimately switched to a different method, which was bulk buying print copies of millions of books, then cutting off the binding of the book and then scanning them at an incredibly fast rate. They literally hired a former executive in charge of the Google Books program, which almost all of our listeners, I'm sure, have at some point encountered. Search results from Google Books. So Anthropic basically acknowledged that what they were doing was illegal. They said, we think we can find a way that would qualify as copyright fair use, which is buying a print copy and then converting it to a digital copy, and then we're going to use that in our training library. And according to the argument that the companies are making, as long as they're not serving the copies of the books in whole or in significant part to the end users of their AI models, it's fair use. Then it's fair use. And so both of these cases really hinge on what is copyright fair use. And that's why these cases are so interesting, because this has really important implications for the future of AI. And so they have a Group of authors, people who write books. In one case, it's a class action lawsuit. That's Meta's case. In the other case, the anthropic, it's just a group of authors who are suing, you know, on behalf of themselves and their own rights. But in both cases, it's probably the two most important rulings thus far on what fair use constitutes in the AI era.
- [11:57]
B
Okay, so what exactly did the judge rule in each case?
- [12:01]
A
Okay, so we've got two cases. And I would say, again, I want to encourage folks to actually go out there and read the orders because I think a lot of the reporting on this has kind of gotten it wrong. Some of the reporting out there is basically saying that, you know, AI training is fair use, using books for AI is fair use. And you can kind of get that conclusion from here. But I think the more accurate telling of the story is that we have two orders that very explicitly disagree with each other. But, you know, the courts have the ability to decide cases in front of them. You know, the powers of the US Judiciary are circumscribed by the Constitution. So the judge who was ruling in the Meta case, which was Vince Chabria, the United States District Court, Northern District of California, he said something that I think is really interesting. This is a very long quote, but I think it's worth quoting at length. Companies are presently racing to develop generative artificial intelligence models, software products that are capable of generating text, images, videos, or sound based on the materials they've previously been trained on. Because the performance of a generative AI model depends on the amount and quality of data it absorbs as part of its training, companies have been unable to resist the temptation to feed copyright protected materials into their models without getting permission from the copyright holders or paying them for the right to use their works for this purpose. This case presents the question whether such conduct is illegal. Although the devil is in the details, in most cases, the answer will likely be yes, meaning, yes, it is illegal. What copyright law cares about above all else is preserving the incentive for human beings to create artistic and scientific works. Therefore, it is generally illegal to copy protected works without permission. And the doctrine of fair use, which provides a defense to certain claims of copyright infringement, typically doesn't apply to copying. That will significantly diminish the ability of copyright holders to make money from their works. Generative AI has the potential to flood the market with endless amounts of images, songs, articles, books, and more. People can prompt generative AI models to produce these outputs using a tiny fraction of the time and creativity that would otherwise be required. So by training generative AI models with copyrighted works, companies are creating something that will, that often will dramatically undermine the market for those works and thus dramatically undermine the incentive for human beings to create things. Things the old fashioned way. So you hear that and you think, wow, Meta must have gotten creamed in this ruling. Actually, no, they won.
- [14:51]
B
Right, and so explain that, explain that.
- [14:54]
A
The reason why they won is that. And the judge is pretty vicious against the plaintiffs in their arguments. He says, quote, this ruling does not stand for the proposition that Meta's use of copyrighted materials to train its language models is lawful. It stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record of support of the right one. So what he's basically saying is, I as a judge am only allowed to rule about the case in front of.
- [15:26]
B
Me that you presented.
- [15:28]
A
Yes, exactly. And in the case of fair use, you know, there are these four things that really matter for whether or not something is illegal. So here are the four factors. The purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used in relation to the copyrighted work as a whole. And then number four, and the most important in this case, the effect of the use upon the potential market for or value of the copyrighted work. And what he's saying is the plaintiffs foolishly made almost all of their case based on those first three factors where the judge does not think there is a grounds to, you know, rule Meta as having done something illegal. And in the fourth case, Meta basically wins by default because they did make an argument on the fourth and the plaintiffs did not. So the judge is basically saying, you know, you morons, if you had come to me with a case around focused on factor number four, the effect of the use on the potential market, I would have ruled in favor of you. But because my powers are circumscribed, I'm only allowed to do the sorts of things that judges are allowed to do in fair use copyright cases. I am literally forced to rule in favor of Meta. So now what's interesting is that the Meta ruling came out a few days after the anthropic ruling. So in the anthropic case, they did advance the fourth argument, the one about the effect on the potential market, and they said that actually this was not a significant factor. So Judge Chabria basically takes Judge Alsup, who's the other judge ruling in the anthropic case, to task. So here's what he here's what he wrote in terms of, you know, attacking his judicial colleague. Speaking of which, in a recent ruling on this topic literally days before, Judge Alsop focused heavily on the transformative nature of generative AI, while brushing aside concerns about the harm it can inflict on the market for the works it's trained on. Such harm would be no different, he reasoned, than the harm caused by using the works for, quote, training school children to write well, which could, quote, result in an explosion of competing works. According to Judge Alsop, this is not the kind of competitive or creative displacement that concerns the Copyright act, however. But when it comes to market effects, using books to teach children to write is not remotely like using books to create a product that a single individual could employ to generate countless competing works with a minuscule fraction of the time and creativity. And it would otherwise take this inapt analogy is not a basis for blowing off the most important factor in the fair use analysis. So this is really interesting. I want to emphasize that when he's using the word transformative, he doesn't mean that, you know, oh my gosh, AI technology is so transformative and amazing. What he means is, are they taking the work, you know, cranking through it and creating a new work, a transformation of the input, or are they just reproducing the input? So most people are not arguing, right, that if you ask Lama, you know, please recite all of the text of Chip War, it will basically refuse to do that. It's not easy to imagine a scenario in which it would do that. But what he's saying is that you now have all the insights from a book, say like Chris Miller's Chip War about the geopolitics of semiconductors.
- [19:01]
B
Also a friend of the podcast, Also.
- [19:03]
A
A friend of the podcast. And maybe you're less likely to buy that book now because rather than buying the book, which will cost you, you know, 25 bucks at Barnes and Noble or Amazon or whatever, you will instead have a conversation with an agent that has internalized all of that hard won research that Chris Miller did. And it's not just that he's going to, you know, that Metta is going to be able to serve that up to you. It's that Meta is going to serve that up to hundreds of millions of people around the world. So he's basically saying that this just a difference in kind, a difference in scale, and it is the creation of a product that could damage the market for, and damage the incentive for humans to create stuff. So if you're like the New York Times, which has very famously sued OpenAI, for example. They're loving the legal analysis that Judge Chabria is putting out, and they're not loving the stuff that they've seen come out of the Anthropic case. I should point out, you know, both of these judges were appointed by Democrats. In Alsop's case, it was Bill Clinton. In Chabir's case, it was Obama. So this is not a partisan kind of breakdown on the issue. This is just a pretty good reflection of where we are. And I think because only binding precedent can come from higher courts, we won't really know what US Law says on this issue definitively until it comes from an appeals court or the Supreme Court or until Congress clarifies, you know, the language in the relevant legislation, the relevant bills on this issue. So it was a. It was a big move. AI companies have reasons to both be encouraged and worried. Copyright holders have reasons to both be encouraged and worried. And a pretty interesting story altogether.
- [20:52]
B
Absolutely. I want to move on. We have a little bit more time left today and talk about the House Select Committee on China. On June 25, the House Select Committee on China held a hearing titled Algorithms and Authoritarians why US AI Must Lead. Witness included three policy experts, among them Anthropic's co founder and head of policy, Jack Clark. Greg, can you start by giving us some context for this hearing? What was it about?
- [21:18]
A
Yeah, so this. I actually had the opportunity to brief a bunch of members of the Select Committee the day before this hearing as they were sort of getting ready. And it's very clear that, unsurprisingly, the Select Committee on China is really interested in AI because they see AI as one of the focal points of US China competition overall. And if you look at some of the language that was coming out of the members of Congress who were leading this hearing, there's some pretty interesting quotes that sort of reflect what the state of opinion is in Congress today. So Chairman Moliner, the Republican from Michigan, opened by saying, quote, today's hearing addresses a defining question for this century. Will the future of artificial intelligence be led by free nations or by authoritarian regimes like the Chinese Communist Party? And I think what's interesting is this time this hearing comes only, you know, a little over a month after a May 8 hearing in the Senate that had executives from OpenAI from the chip design company AMD, from Core Weave, a big cloud computing provider, and Microsoft, also a big cloud computing provider and a big AI company in its own right. This time they brought in Dr. Thomas Manken, who's from the center for Strategic and Budgetary Assessments, a government backed think tank. Mark Beale, my former colleague in the DoD, who's now with the AI Policy Network, and Jack Clark, who's at Anthropic, one of the co founders over there. And what's so interesting is there was a big contrast in the focus of the AI China challenge between this hearing and that last hearing, because both agree that, you know, competition with China is really important for the United States in AI. But in this hearing, we actually tallied up the number of times that people said artificial general intelligence or artificial superintelligence. And it was mentioned no less than 15 times in this House hearing. It came up zero times in May's Senate hearing. And a lot of the times it was being mentioned, it was being mentioned by members of Congress, so the United States government. You know, I remember when I was in the dod, it was, you know, something, if you were to talk about AI being smarter than humans, if you were to talk about AI being way smarter than humans, something like artificial superintelligence, that was kind of seen as a fringe opinion in policy circles, the kind of thing it was almost vaguely impolite to say in public. Now you have members of Congress who are saying stuff like this. You know, this comes from Rep. Torres, a Democrat of New York, quote, In the 20th century, the United States and Nazi Germany were locked in a high stakes race to build the first atomic bomb in the 21st century. The United States and China are competing in a new strategic arms race, the race for artificial superintelligence. And the first country to reach ASI will likely emerge as the superpower of the 21st century. The superpower who will set the rules for the rest of the world. That was just not something that you could say in public when I was in DoD and reflects just how much this debate has changed and now is focused on AGI and asi.
- [24:42]
B
All right, so what were some of the things that really stood out to you from the hearing?
- [24:47]
A
Well, what's interesting is while there was this focus on competition with China and what we need to do to compete with China, there was also a related conversation about safety. There was a lovely point made by Jack Clark of Anthropic where he said, America can win the race to build powerful AI and winning this race is a necessary but not sufficient achievement. We have to get safety. Right. So what he's basically saying is, and this is, you know, a guy who works at a big AI company and he's saying, America faces a threat from China in AI. But America also faces a threat from unsafe AI, and he's trying to get Congress to not throw out a focus on safety as it continues a focus on competition with China. And so Representative Krishna Murthy, who is the ranking member, actually announced that he's going to be introducing at some point in the future. You know, he. This is the first time he talked about it in public. Working on a new bill, the AGI Safety act, which is going to require AGI to be aligned with human values and require it to comply with laws that apply to humans. So that's kind of interesting is at the same time, people are really trying to push this China narrative. There's still momentum, including from the China Hawk community about addressing the threat of AI, which, you know, has really been a prominent point of discussion ever since that letter was signed in the middle of 2023, connecting AI to the risk of human extinction.
- [26:22]
B
Greg, there was another one, no Adversarial AI Act. So there were two new AI bills. What does the Adversarial AI act do?
- [26:31]
A
Yeah, thanks for mentioning this, because I almost forgot about it. The Adversarial AI act is about blocking the US Government from using AI models that come from China or Russia. So you don't want. And this is, you know, also something that Krishnamurthy is backing, but it's bipartisan. And this is, you know, preventing federal government agencies from using stuff like Deep Seek, which, you know, as. As I think we're about to talk about. There's a lot of concern that that is a security threat in and of itself that, you know, if you're using Deep seq, that could either be sort of like a sleeper cell that could activate cybersecurity vulnerabilities, or, you know, as AI becomes more agentic in the future, you know, the AI agent itself could be more threatening. And so that's just to block agencies from using these systems unless they're, like, evaluating them, you know, for, like, benchmarking them against US models or for the security vulnerabilities, et cetera. I have to assume some version of that will pass, whether it comes from Congress or whether it comes in the form of an executive order. It's really tough for me to envision a scenario in which the US Government is comfortable using Chinese or Russian AI models.
- [27:44]
B
Speaking of China, we can't talk about China without talking about Deep Seq these days. So I want to discuss two recent developments relating to China's Deep seq. First, US Export controls have repeatedly delayed the release of its new model R2. And second, a US official said that Deep Sea aids the Chinese military and, and evades export controls. Let's start with the delay. How are US Export controls affecting Deep Seq's timeline?
- [28:10]
A
Yeah, so as, as folks will remember, you know, the, the CEO of Deepseek had said in a 2023 interview and a 2024 interview that export controls were basically the biggest challenge that his company was facing, that compute limitations were the biggest challenge the company was facing. Well, R1 was the model that, you know, really took the world by storm, resulted in a temporary, you know, big dip in Nvidia stock price, although it is more than fully recovered since as people were sort of wondering, okay, can, you know, China innovate its way around these export controls? Well, what we're finding out now is that according to reporting, some really good reporting in the information, which is interviewing folks in Deep Seq or in its supply chain or who are partners of the company, they're not satisfied with the performance of R2. One of the reasons why they're not satisfied with the performance of R2 is that they're struggling to come up with a version of the model that's going to be deployable given the new shortage in China of these Nvidia H20 chips, which the Trump administration banned. It also, according to reporting from Reuters, citing, you know, a conversation with a senior State Department official who is in this interview anonymous, but said that they're tracking active attempts by Deep Sea to smuggle chips, that they're tracking Deep Sea collaboration with Chinese military and intelligence services. And overall, I think the export controls, you know, are having this impact on Deep Seq. They're not able to scale the way that they want to. And I think Jin Yang, who is the Asia Bureau Chief at the Information, who posted this on X, I think put it quite well, quote, Deep Seq's models are so completely optimized for Nvidia hardware and software that running them on Chinese chips will make them less efficient. And so the point basically is here that even though Deep Seq is a genuinely innovative company, even though they have come up with ways to squeeze more AI performance out of weaker and fewer chips, they still want Nvidia weaker chips, namely those H20 chips that are now banned. And they're running into a lot of challenges with compute limitations. I think, you know, over a 3, 5, 10 year time frame, Huawei will be a more relevant competitor in this story. But at least for the next few years, they're stuck with a pretty lousy chip design in the case of the 910B, they're stuck with a lousy software ecosystem and they're stuck with pretty significant production challenges, which also relates to the export controls on semiconductor manufacturing equipment. So all of this, I think, is to say that the export controls, while a highly imperfect tool, while the impact that they will have cannot last more than a decade or two at the absolute most. Right. In this case, they are indeed having the intended impact, just not to the same extent that was originally desired.
- [31:17]
B
Craig, maybe most importantly, I have to ask you, what do we know about Deep Seat's role in supporting China's military and intelligence apparatus?
- [31:25]
A
Sure. So the report from Reuters is talking about Deep sea collaboration with Chinese surveillance apparatus, which is basically China. If they want to have access to all of your text messages, all of your emails, the Internet giants know that they have no ability to tell the Chinese government no. So the idea that that would now extend to chat logs with Deep Seq, at least in the case when it's going over via API, as opposed to something that's running locally on your own hardware, makes total sense to me. Right. Deepseek would have no ability to say no to the Chinese government when they demand that sort of information. In terms of Deep Seq actively aiding the military and intelligence services, I don't believe there's any direct reporting on this other quoting this US official, but it's a pretty obvious state of affairs because China is pushing the adoption of Deep Seq AI models in all federal and state and local government agencies. I mean, there's like a bonanza of the Chinese government saying adopt Deep Seq. Adopt Deep Seq. This is our local AI champion. You know, you need in your quarterly reports to explain what you're doing to effectively adopt Deep Seq. So of course that exact same initiative is going to affect the defense and intelligence services. You know just how effective they're being when they adopt AI. Maybe they're using it for incredible use cases, maybe they're using it for silly use cases. You know, that kind of information is not out there in the public reporting or in the disclosures from U.S. and Defense Intelligence community kind of thing. But it is a very easy connecting of the dots to say that an order that affects every Chinese government agency is going to affect defense and intelligence agencies.
- [33:13]
B
Greg, as usual, thanks for the tremendous insight. These are all fascinating issues and really appreciate it. To all our listeners, Happy 4th of July and we will be back after the holiday. Thanks so much.
- [33:25]
A
Thanks, Andrew. Thanks for listening to this week's episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps spread the word. This podcast was produced by Sarah Baker, Isaac Goldston and Sadie McCullough. See you next time.